Is Avalon-M1-Mini state of the art?Pending benchmarks.Announced Oct 23, 2026; no independent results yet.
General Reasoning's research model trained to work as an agent organisation: copies of one policy act as a CEO and hired specialists sharing a message board and workspace, rewarded as a team (reinforcement learning from collective rewards, RLCR). Post-trained from NVIDIA's Nemotron 3 Super base. Training is still running; results so far are a preview, and weights are planned but not out.
The details so far
- Lab
- General Reasoning
- Announced
- Oct 23, 2026
- API
- Not yet
- Open weights
- Announced, not yet published
- Parameters
- 120B total / 12B active
- Base model
- NVIDIA Nemotron 3 Super pretrained base (120B-A12B)
- Recipe
- Reinforcement learning from collective rewards (RLCR): every agent in the organisation gets the same collective advantage
- Training data
- Small SFT set of 38,374 examples, then RL over 147 environments with more than 2 million tasks
- Status
- Still training; the technical report is marked "under revision"
- Weights
- Not released; the lab's roadmap plans to release the model weights, harness and open environments
- API
- None announced
- Paper
- Technical report on gr.inc; final version planned for arXiv
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: in the lab's organisational evaluations, Avalon-M1-Mini outperforms Nemotron 3 Super, post-trained from the same base without on-policy distillation, on every benchmark shown.
Lab claim: against Inkling Small, comparing each model's better configuration, Avalon-M1-Mini wins 6 of 18 benchmarks and ties 3 (all run on the lab's own infrastructure).
Lab claim: working as an organisation instead of alone raises IMO-ProofBench-Advanced from 29.8% to 37.5% and FrontierScience from 55.9% to 63.5%.
Sources
LABGeneral Reasoning announcement on Xx.com/GenReasoning/status/2113590771722952908 · Oct 23, 2026LABAvalon-M1 release pagegr.inc/releases/avalon-m1 · Oct 23, 2026PAPERAvalon-M1: Reinforcement Learning from Collective Rewards (technical report)gr.inc/avalonm1.pdf · Oct 23, 2026LABRoss Taylor (General Reasoning) post on the base modelx.com/rosstaylor90/status/2113616941646573716 · Oct 23, 2026Tracking ends Jan 21, 2027 unless the API or weights become available.